基于多模态数据的PET图像无监督去噪方法研究

正电子发射断层扫描（PET）图像存在低信噪比的问题，对临床诊断有很大的影响。近年来，许多基于深度学习的医学图像去噪方法被提出并取得了良好的效果，但这些方法通常需要低质量-高质量图像对进行监督训练。在PET的临床应用中，存在对放射性示踪剂剂量安全性的考虑和扫描采集时间的限制，通常很难获得大量的高质量PET图像。我们注意到由于近年来PET/CT和PET/MR的普及，临床上更易获得的是噪声PET图像以及与其相对应的CT或MR图像（低质量-解剖结构图像对）。本研究在考虑临床实践中PET数据特性的前提下，提出利用解剖结构图像替代高质量PET图像，建立基于低质量-解剖结构图像对的无监督深度学习网络，并且在系统研究基于示踪剂动力学模型的PET参数图像估计方法的基础上，融合参数图像估计模型和无监督深度学习去噪方法，从而实现PET静态图像和参数图像去噪，为临床诊断治疗和药物研究提供高质量的PET图像。

The signal-to-noise ratio of Positron emission tomography (PET) image is low, which has a great impact on clinical diagnosis. Recently, many deep-learning-based medical image denoising methods have been proposed and achieved good performance, but these methods usually need low- and high-quality image pairs for supervised training. However, in real clinical practice, it is difficult to obtain a large number of high-quality PET images due to the dose concern and the scanning time limitation. We noticed that due to the development of PET/CT and PET/MR, it is easier to obtain noisy PET images and corresponding CT or MR images (low-quality and anatomical image pairs). In this study, considering the characteristics of PET data, by using the anatomical image instead of the high-quality PET image, an unsupervised deep learning network based on low-quality and anatomical image pairs will be established. Then, after a systematic study of PET parameter image estimation methods based on tracer kinetic modeling, we will combine the parameter image estimation model and the unsupervised deep learning denoising network together. Our goal is to achieve denoising of both PET static images and parameter images and provide high-quality PET images for clinical diagnosis, treatment, and drug research.